Image: GitHub
Why it mattersDevelopers who run more than one coding agent stop editing four config files by hand to try a different model on the same task, and can point every agent at one signed-in subscription.
A developer with Claude Code, Codex and Cursor open on the same laptop edits three separate config files to try a cheaper model on the same coding task. A new tool released yesterday moves the choice into a single menu bar dropdown.
yetone opened the magpie repository at 16:22 UTC on 23 September 2026. It reached 213 stars and 8 forks in the first day, is MIT-licensed, and lives at github.com/yetone/magpie. Read from the GitHub API at 04:10 UTC on 24 September.
What it does
magpie is one screen that lists each AI coding agent installed on the machine and the model it is set to. The panel drops from the menu bar, or opens as a normal window, or runs in the terminal. A click on a model name shows a picker; a save writes the choice back to the agent's own config file.
The README lists nine agents it manages: Claude Code, Codex, Gemini CLI, OpenCode, Pi, Goose, Cursor CLI, Copilot CLI and Crush. Each has its own file path and its own idea of what a model name looks like. yetone says magpie only touches the one key that changes, so comments, ordering and indentation in settings.json, config.toml, opencode.jsonc or config.yaml survive intact, and writes are atomic.
The gateway is the interesting part
Every agent's picker points at the same catalog because magpie runs a local gateway on 127.0.0.1:3425. The gateway speaks the OpenAI chat completions, OpenAI Responses, Anthropic Messages and Google Gemini APIs, and translates between them on the way to whichever vendor serves the picked model. Streaming and tool calls are included.
yetone lists Anthropic, OpenAI, DeepSeek, Kimi, GLM, MiniMax, Qwen, Mistral, Groq, xAI, OpenRouter, Together, Fireworks, Ollama and LM Studio among the presets, and says a custom vendor needs a name and a base URL. magpie asks each vendor which models it serves and offers exactly those, rather than shipping a hardcoded list, so a model released this morning shows up on the next refresh.
One subscription, every agent
An agent already signed in to Claude Code, ChatGPT or Copilot is treated as a provider in itself. magpie reads the agent's own credentials each time and rotates its tokens the way that agent would, and every other agent on the machine can pick that subscription's models through the gateway. yetone says a Claude subscription is a special case, because Anthropic classifies another agent's system prompt as third-party traffic even over the same OAuth login. magpie drives the real local claude binary in that case and bridges the caller's tools in over MCP, so the request looks like a Claude Code turn to Anthropic.
Anyone running more than one coding agent hits the same problem: three or four config files with different shapes, three or four vendor keys, and a different model picker in each tool. magpie centralises the picker, and by putting a gateway underneath it removes the need for each agent to hold vendor keys at all. 213 stars in the first day is early, so this is a project to watch and the shape it will settle into is still open. The design points at a working bet: the coding-agent stack is heading toward one shared model catalogue and one shared account layer, so the vendor a developer stays with becomes whoever owns the subscription.
Source
- yetone/magpie on GitHub, MIT-licensed, primary source and README verified 24 September 2026. Stars, forks and creation date read from
api.github.com/repos/yetone/magpieat 04:10 UTC.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
